babl-fish-reference: do some conversions in single precision float

babl-fish-reference: do some conversions in single precision float

This permits these conversions to go through floating point instead of double
precision when non of the involved format have more than 32bit precision,
this might be determinetal to u32 but is right for u8 and u16 which we rely
most on.

We also disable the missing fast path messages, since now many references;
through internal shortcuts are faster than some long chains, perhaps some
of these reference fast paths should be split out as new categories of fishes.